fix: https://github.com/triton-inference-server/server/issues/8537

[aleksn7]

Fix the race between the done callback async_event_future_done_callback and the background_futures_.attr("add")

Pull request for the issue described here:
triton-inference-server/server#8537

[whoisj]

Thank you for your contribution. @aleksn7, have you signed a CLA? If not, please see Contributing / CLA.

[Copilot]

Pull request overview

This PR addresses a race condition in the Python backend stub’s async coroutine handling for decoupled execution, ensuring background futures are tracked before their completion callback can run.

Changes:

• Reorders background_futures_.add() to occur before add_done_callback() in Stub::RunCoroutine.
• Adds an in-code comment referencing the upstream issue for context.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[aleksn7]

This should never happen.

[aleksn7]

Thank you for your contribution. @aleksn7, have you signed a CLA? If not, please see Contributing / CLA.

Hello! I've just signed the CLA

[whoisj]

LGTM. CLA received. @yinggeh can you take a quick look as well? Thanks.

[yinggeh]

Pipeline ID: 51138619

[yinggeh]

LGTM. However, this change seems to introduce test failures. It might be noise but still unclear. We will merge the PR once it's clear. cc @whoisj

[aleksn7]

LGTM. However, this change seems to introduce test failures. It might be noise but still unclear. We will merge the PR once it's clear. cc @whoisj

Hi! Do you have any updates on this?

[whoisj]

@aleksn7, apologies for slow reply as we've been off work for the past five days. I'm looking into the test failures. Looks like they might be caused by something unrelated to this PR, but they do make validating the PR difficult. We'll return to this PR once the test failures have been resolved. Again, apologies for the delay.